Tag
This paper explores how transformer models handle functional words like pronouns and adverbs, analyzing their embeddings and generalization capabilities. It finds that training on mixed lexicalized and functional sentences helps models learn shared syntactic-semantic structures.